The verdict in three sentences
An internal AI knowledge assistant connected to your procedures, contracts, technical sheets and minutes answers in seconds with the source cited, while your teams currently spend up to 2 hours a day searching. A RAG (retrieval-augmented generation) project for a mid-sized company costs EUR 25,000 to 70,000, with usage costs of EUR 300 to 2,000 a month depending on user count. Two requirements are non-negotiable: EU hosting and enforcement of existing access rights, so an intern never gets an answer drawn from a confidential HR document.
Three project levels priced
| Level | Scope | 2026 budget | Timeline |
|---|---|---|---|
| Pilot | 1 department, 5,000 documents, SharePoint or Drive, 50 users | EUR 25,000 to 35,000 | 6 to 8 weeks |
| Company rollout | 40,000 documents, 3 to 5 sources, synced permissions, 500 users | EUR 40,000 to 55,000 | 8 to 10 weeks |
| Advanced | Several languages, Teams integration, business actions, usage dashboards | EUR 55,000 to 70,000 | 10 to 12 weeks |
| Monthly usage cost (model and hosting) | Depending on question volume | EUR 300 to 2,000 per month | Ongoing |
| Maintenance and improvement | Re-indexing, answer evaluation | EUR 500 to 1,500 per month | Ongoing |
Usage cost depends mainly on question volume: an active user asks 8 to 15 questions a day on average, a few cents each with a recent model hosted in Europe.
Where a RAG budget goes
The language model itself is the cheapest part. Most of the work goes into document preparation and security.
| Work package | Share of budget | Watch-out |
|---|---|---|
| Source inventory and clean-up | 10 to 15% | Obsolete versions, duplicates, unreadable scans |
| Connectors (SharePoint, Drive, DMS, intranet) | 15 to 20% | Incremental nightly sync |
| Chunking, indexing and hybrid search | 20 to 25% | Tables, drawings, scanned PDFs |
| Access rights management | 15 to 20% | Inheriting Active Directory or Entra ID groups |
| Interface (web, Teams) and citations | 10 to 15% | Direct link to the source paragraph |
| Evaluation and tuning | 10 to 15% | 200 reference questions, correct answer rate |
A serious project measures a correct answer rate on a question set validated by the business. Targeting 85 to 90% at launch is realistic, with an "I don't know" answer rather than an invention when the source is missing.
Compliance: what your DPO will ask
The EU AI Act classes this kind of internal assistant as limited risk, but GDPR applies in full as soon as documents contain personal data. Require EU hosting, no reuse of your data to train the model, question logs retained for 6 to 12 months, and an impact assessment if HR or customer files are indexed. In Germany, the works council (Betriebsrat) must usually be consulted when usage logs could monitor employees. In 2026, model providers offer hosting in Germany or France, or open models you can install on your own cloud.
Mini case study
Nathalie, CIO of a 900-person engineering firm in Berlin, manages 40,000 documents split between SharePoint and a legacy DMS. An internal survey shows that 200 engineers spend an average of 2 h a day searching.
- Rollout project: EUR 48,000, usage and maintenance EUR 1,800 per month
- Prudent assumption: 20% of search time saved, only half converted into productive work, i.e. 12 min per engineer per day
- Value: 200 x 12 min x 20 days = 800 h per month x EUR 55 = EUR 44,000 per month
- Payback: 48,000 / (44,000 - 1,800) = about 1.1 months
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
Even dividing that gain by four, the project pays back in under 5 months. That is why Nathalie started with a EUR 30,000 pilot in the design office to measure the real gain.
FAQ
Why not just use Copilot or a consumer chatbot?
Copilot for Microsoft 365 costs about EUR 30 per user per month, i.e. EUR 180,000 a year for 500 people, and covers third-party DMS poorly. A dedicated RAG makes sense beyond 150 to 200 users or with non-Microsoft sources.
Can the assistant make up an answer?
The risk exists but drops sharply when the model must cite its sources and answer "I don't know" when no relevant passage is found. A 5 to 10% error rate at launch is common, then falls with tuning.
How long does indexing 40,000 documents take?
Initial indexing takes 1 to 3 days of compute. Updates then run nightly, reprocessing only modified documents.
Are access rights really enforced?
Yes if filtering happens before retrieval: each indexed passage carries the list of authorised groups, synced from Entra ID or Active Directory.
Can we use it inside Teams?
Yes, a Teams integration usually adds EUR 3,000 to 6,000 and strongly boosts adoption.
Let's scope your project. Describe your document sources, user count and hosting constraints: we will price a pilot or rollout between EUR 25,000 and 70,000, delivered in 8 to 12 weeks. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
